Shaleen Jain
I'm a computer vision engineer with a strong foundation in classical image processing and a passion for modern AI. My journey began at BARC, where I tackled scientific imaging challenges, and evolved at Applied Materials, where I built deep learning-based solutions for semiconductor manufacturing. I specialize in fusing traditional vision techniques with deep learning to create scalable, interpretable, and high-performance systems. I thrive at the intersection of research and real-world application—bringing AI from concept to product.
Beyond the lab and code, I find balance and inspiration in the outdoors and the arts. Trekking fuels my sense of exploration, much like research does. Reading and writing in Hindi keep me grounded in language and expression, while cooking—both a science and an art—mirrors my love for experimentation and creativity. Whether it's crafting a model or a meal, I enjoy the process of building something meaningful from the ground up.
Session
Python is often criticized for its Global Interpreter Lock (GIL), which is seen as a bottleneck for high-performance computing. However, this talk showcases how Python, when used with right design principle, can deliver 40x throughput improvement w.r.t. baseline implementation in a real-world image segmentation post-processing pipeline. We present a generic rectangle partitioning algorithm that converts irregular segments into precise, non-overlapping rectangles.
By leveraging multiprocessing, ProcessPoolExecutor, and Numba JIT compilation with shared memory and inter-process communication (IPC), we overcome GIL limitations and scale efficiently across cores. This session will walk through the architectural decisions, performance bottlenecks, and Pythonic optimizations that made this possible—demonstrating that Python, with the right tools and mindset, can be both elegant and fast.